mmsegmentation-distiller

mmsegmentation-distiller copied to clipboard

ChannelWiseDivergence unused variable

1

In mmseg.distillation.losses.cwd.ChannelWiseDivergence the student softmax output is not used (also self.name): `softmax_pred_S = F.softmax(preds_S.view(-1,W*H)/self.tau, dim=1)` Intended or a bug?

RedHeadM

Question about teacher model performance in the paper "channel wise distillation for semantic segmentation"

1

Hi, I happen to found that the results of PSPnet-r101 on Pascal VOC（mIoU=78.52, mAcc=79.57）is exactly the same as the official model provided by mmsegmentation. 1、Have you tested the pretrained model?...

ofishhaha

when inference at 8000 iter, error in 'SegmentationDistiller' object has no attribute 'test_cfg'

3


Wangman1

Is the Experimental settings same as the paper？

2

Thank you for your excellent project! As I noticed, the experimental settings, like image crop size, batch size, training iterations are not the same as the distillation paper "channel wise...

ofishhaha